提交 68399ab9 编写于 作者: S superjom

Merge remote-tracking branch 'origin/rnn-backward-python' into feature/recurrent_op_backward_fix

......@@ -80,7 +80,6 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope) const {
// Now all variables in scope must be created outside of op.
PADDLE_ENFORCE_NOT_NULL(stepnet_);
PADDLE_ENFORCE(!(*stepnet_)->Outputs().empty(), "stepnet_ op has no outputs");
PADDLE_ENFORCE(!(*stepnet_)->Outputs().empty(), "net_op has no outputs");
if (seq_len_ > step_scopes->size()) {
for (size_t i = step_scopes->size(); i < seq_len_; ++i) {
......
......@@ -311,6 +311,15 @@ All parameter, weight, gradient are variables in Paddle.
self.set_falsenet(net.Clone());
});
rnn.def("backward",
[](const operators::RecurrentOp &forwardOp,
const std::unordered_set<std::string> &no_grad_vars) {
const auto &op = *static_cast<const OperatorBase *>(&forwardOp);
return Backward(op, no_grad_vars);
});
ExposeOperator(rnn);
m.def("unique_integer", UniqueIntegerGenerator);
m.def("is_compile_gpu", IsCompileGPU);
......
......@@ -3,6 +3,7 @@ import paddle.v2.framework.core as core
import unittest
import numpy as np
from paddle.v2.framework.op import Operator, RecurrentOp
from gradient_checker import GradientChecker
def py_sigmoid(x):
......@@ -68,7 +69,7 @@ def create_tensor(scope, name, shape, np_data):
return tensor
class TestRecurrentOp(unittest.TestCase):
class RecurrentOpTest(unittest.TestCase):
'''
Test RNNOp
......@@ -160,5 +161,42 @@ class TestRecurrentOp(unittest.TestCase):
self.assertEqual(pd_output.shape, py_output.shape)
class RecurrentGradientOpTest(unittest.TestCase):
def create_forward_op(self):
self.forward_op = RecurrentOp(
# inputs
inlinks=["x"],
boot_memories=["h_boot"],
step_net="stepnet",
# outputs
outlinks=["h"],
step_scopes="step_scopes",
# attributes
inlink_alias=["x@alias"],
outlink_alias=["h@alias"],
pre_memories=["h@pre"],
memories=["h@alias"])
# create a stepnet for RNN
stepnet = core.Net.create()
x_fc_op = Operator("mul", X="x@alias", Y="W", Out="Wx")
h_fc_op = Operator("mul", X="h@pre", Y="U", Out="Uh")
sum_op = Operator("add_two", X="Wx", Y="Uh", Out="sum")
sig_op = Operator("sigmoid", X="sum", Y="h@alias")
for op in [x_fc_op, h_fc_op, sum_op, sig_op]:
stepnet.add_op(op)
stepnet.complete_add_op(True)
self.forward_op.set_stepnet(stepnet)
def create_gradient_op(self):
a = set()
backward_op = core.RecurrentOp.backward(self.forward_op, a)
def test_grad(self):
self.create_forward_op()
self.create_gradient_op()
if __name__ == '__main__':
unittest.main()
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册